Scholarlink

<p>Realized volatility exhibits complex non-linear and non-stationary characteristics, making its accurate prediction challenging yet essential for effective financial risk management and investment decision-making. This study proposes a hybrid decomposition-ensemble approach to predict the realized volatility of four stocks: 600519.SS, 601398.SS, 601857.SS, and 601988.SS. Variational Mode Decomposition (VMD), optimized by Signal-to-Noise Ratio (SNR), is employed to decompose the volatility time series into Intrinsic Mode Functions (IMFs) and residuals, effectively capturing the intricate patterns inherent in the data. Sample entropy and t-tests are utilized to classify these components into high-frequency and low-frequency segments. The low-frequency components, representing long-term trends, are directly fed into various Deep Neural Network (DNN) architectures, including Simple Recurrent Neural Network (SRNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional GRU (BiGRU), and Bidirectional LSTM (BiLSTM). High-frequency components, capturing short-term fluctuations, are modeled using Generalized Autoregressive Conditional Heteroskedasticity (GARCH) to extract conditional volatility, which is subsequently inputted into the DNN models. The final realized volatility prediction is obtained by linearly superimposing the outputs from all components, forming a comprehensive ensemble model. The proposed hybrid models are benchmarked against traditional standalone DNN models using evaluation metrics such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and R-squared (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(R^2\)</EquationSource> </InlineEquation>). Empirical results demonstrate that the hybrid VMD-GARCH-DNN models significantly outperform traditional models across all metrics, effectively capturing both linear and non-linear dynamics of realized volatility. The models are further evaluated through multi-step forecasting, with 10-day and 22-day ahead forecasts, representing typical investment horizons. The results highlight the superior performance of the proposed models across both short and longer-term horizons, offering a robust and accurate framework for volatility forecasting. This approach provides valuable insights for financial analysts and practitioners in managing risk and making informed investment decisions.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Decomposition-Ensemble Approach for Realized Volatility Prediction

  • 期刊论文

  • 发表日期：

    2025年6月18日

查看全文
  • John Kamwele Mutinda,
  • Li Yong

摘要

Realized volatility exhibits complex non-linear and non-stationary characteristics, making its accurate prediction challenging yet essential for effective financial risk management and investment decision-making. This study proposes a hybrid decomposition-ensemble approach to predict the realized volatility of four stocks: 600519.SS, 601398.SS, 601857.SS, and 601988.SS. Variational Mode Decomposition (VMD), optimized by Signal-to-Noise Ratio (SNR), is employed to decompose the volatility time series into Intrinsic Mode Functions (IMFs) and residuals, effectively capturing the intricate patterns inherent in the data. Sample entropy and t-tests are utilized to classify these components into high-frequency and low-frequency segments. The low-frequency components, representing long-term trends, are directly fed into various Deep Neural Network (DNN) architectures, including Simple Recurrent Neural Network (SRNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Bidirectional GRU (BiGRU), and Bidirectional LSTM (BiLSTM). High-frequency components, capturing short-term fluctuations, are modeled using Generalized Autoregressive Conditional Heteroskedasticity (GARCH) to extract conditional volatility, which is subsequently inputted into the DNN models. The final realized volatility prediction is obtained by linearly superimposing the outputs from all components, forming a comprehensive ensemble model. The proposed hybrid models are benchmarked against traditional standalone DNN models using evaluation metrics such as Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Percentage Error (MAPE), and R-squared ( \(R^2\) ). Empirical results demonstrate that the hybrid VMD-GARCH-DNN models significantly outperform traditional models across all metrics, effectively capturing both linear and non-linear dynamics of realized volatility. The models are further evaluated through multi-step forecasting, with 10-day and 22-day ahead forecasts, representing typical investment horizons. The results highlight the superior performance of the proposed models across both short and longer-term horizons, offering a robust and accurate framework for volatility forecasting. This approach provides valuable insights for financial analysts and practitioners in managing risk and making informed investment decisions.